2026-06-16

Kernel benchmark — 2026-06-16

Raw margine-bench-kernel.sh results and the generated comparison for Margine's CachyOS/BORE kernel vs the stock Fedora kernel, plus the published chart.

Conditions

• Hardware: Framework Laptop 13 (AMD Ryzen 7040 Series — Ryzen 5 7640U / Radeon 760M, 12 threads), on AC. The exact board generation is inferable from the CPU.
• Governor: performance. sched_ext (scx) schedulers: off (stock BORE).
• Method: the same margine-bench-kernel.sh on the same laptop, switching ostree deployments — Margine vs ghcr.io/ublue-os/bluefin-dx:stable — so the only variable is the kernel. Benchmark tools run in a throwaway Fedora distrobox (the host has no dnf), so they exercise the real booted kernel.
• Kernels: Margine 7.0.12-cachyos1.fc44, Bluefin/Fedora 7.0.8-200.fc44 (each distro's current stable kernel — different trees, can't be version-matched).
• Thermal control: 4 runs per OS at varied start temperatures, so the median start temp matches (Margine ~51–54 °C, Bluefin ~52 °C → comparable). Every run ended at the CPU's ~100 °C thermal limit, so both kernels throttle equally; absolute numbers are therefore conservative, the relative gap is fair.

Files

• margine-cachyos-N.json, bluefin-dx-N.json — raw per-run results (parsed metrics + kernel/CPU/governor identity + start/end temperature).
• perf-kernel.svg / perf-kernel.md — the published comparison (median).

Published median (and one excluded run)

margine-cachyos-1.json is excluded from the published median: it was taken on an older candidate deployment in a separate session and is a 2× outlier on the context-switch (perf bench sched pipe) metric — its value (~7.7 µs) matches Bluefin's while the other three Margine runs agree at ~4.4 µs. It is kept here for transparency. The median absorbs it either way; dropping it just tightens the set.

Regenerate the published chart (run from this directory):

../../margine-bench-compare.py \
  margine-cachyos-2.json margine-cachyos-3.json margine-cachyos-4.json \
  bluefin-dx-*.json \
  --title "Margine kernel vs stock Fedora — same laptop, median of runs" \
  --out-prefix perf-kernel

Headline (median, this hardware)

Metric	Margine (CachyOS/BORE)	Δ vs stock Fedora
Context-switch latency / rate	4.45 µs/op · 224k ops/s	~1.8× faster
Thread throughput (events)	88k	+54%
Thread latency (avg)	4.1 ms	55% lower
Wakeup latency (p50)	2.08 ms	43% lower
Sched-messaging	2.99 s	9% faster
Tail latency (p95 / p99)	—	~10% higher (BORE trades worst-case for common-case)